[rahulchhabria]

DESCRIBE YOUR PR

Add an opt-in Fin Agent API pilot that preserves the docs Ask AI entry points with a first-party streaming dialog, sanitized answers, citations, follow-up questions, and transcript copying. It loads neither Intercom Messenger nor new browser storage/cookies. Kapa remains the default until explicitly switched.

A protected Vercel preview can now run with the Fin API key, provider/enable flags, and its fixed origin, without Redis or another secret. Signed conversation tokens stay in page memory and carry follow-ups across serverless instances. They expire after one hour and are bound to the configured origin. Same-origin requests preserve an existing preview-protection login cookie; Ask AI does not create one.

Redis is optional for evaluation and required for production. Adding it enables shared locks/budgets, stored helpfulness feedback, signed-webhook evidence capture, authenticated event export, and optional direct support handoff. Without it, those controls are hidden or unavailable. Stateless tokens are replayable, so this evaluation mode has no hard shared spending limit and must stay restricted to testers. A configured Redis failure never silently switches modes. The internal docs-writing/PR worker and human-reply ingestion remain future work.

The pilot guide separates the minimal preview setup from later support and feedback integration. Fin API access and SSE/streaming must be enabled for live answers.

Validation: 36 targeted tests, TypeScript, ESLint, and formatting passed locally. Browser checks with simulated Fin responses passed for desktop/mobile in both modes, including token continuation, capability controls, sanitization, copying, focus, and preserving existing preview authentication without creating cookies or browser storage. GitHub Actions passed all 537 tests across 46 files; lint, security checks, Warden, and Bugbot passed. Both Vercel preview builds passed. Live Fin responses and actual support delivery have not yet been verified.

IS YOUR CHANGE URGENT?

• Urgent deadline (GA date, etc.): YYYY-MM-DD
• Other deadline: YYYY-MM-DD
• No deadline: Not urgent, can wait up to 1 week+

PRE-MERGE CHECKLIST

• Checked Vercel preview for correctness, including links
• PR was reviewed and approved by any necessary SMEs (subject matter experts)
• PR was reviewed and approved by a member of the Sentry docs team

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
develop-docs	Ready	Preview	Oct 24, 2026 5:03am UTC
sentry-docs	Ready	Preview	Oct 24, 2026 5:03am UTC

[cursor]

The plan checks Ask AI reachability, search volume, and server failures. An issue escalates when Ask AI traffic collapses or search errors rise. The Fin dialog stays off until the deployment turns the Fin provider on.

Services: sentry-docs, develop-docs.

Mention @change-monitor in a comment to update the plan.

Plan

What changed

Ask AI buttons now open through a first-party provider. The default path still opens the current Kapa widget. Kapa stays loaded when ASK_AI_PROVIDER is not fin.

The Fin dialog is reachable when ASK_AI_PROVIDER is fin and FIN_PILOT_ENABLED is true. That dialog posts /api/ask-ai/ and streams sanitized answers. Support handoff stays off unless FIN_HANDOFF_ENABLED is true.

sentry-docs and develop-docs must deploy before this change is live. Those Vercel projects share app/ and src/. Catalog ids match that deploy map.

Risk

A provider crash can blank every page. The provider wraps the root layout. Broken click handlers can stop Ask AI or Algolia search. Turning Fin on too soon removes Kapa and can leave Ask AI with failing streams. Public scanners can hit the new routes and receive 503 while the gate stays off.

Intended effect

Confirm that Kapa widget requests continue while Fin stays off. Confirm that /api/ask-ai/ returns 200 streams only after Fin is on. Absent for Kapa is widget volume near zero while Kapa should still load. Absent for Fin is zero 200 responses after Fin is on. Zero /api/ask-ai/ traffic while Fin stays off is expected and does not confirm Fin.

Signal	Baseline	Window	Rule	Source
Kapa widget spans on docs	13023 spans in 24h. 2977 spans in the prior 24h. Band 2977-13023.	2026-10-18T03:25:53Z to 2026-10-19T03:25:53Z and 2026-10-17T03:25:53Z to 2026-10-18T03:25:53Z	While Kapa is default, hold in or above this band. A fall near 0 is absent. After Fin is on, expect a fall toward 0.	Sentry org sentry project docs. Query environment:production (http.url:*widget.kapa.ai* OR http.url:*kapa-widget-proxy*).
Kapa widget spans on develop-docs	287 spans in 24h. 300 spans in the prior 24h. Band 287-300.	Same windows	Same rule on this project.	Sentry org sentry project develop-docs. Same query.
/api/ask-ai/ HTTP spans	0 spans. This route is new.	2026-10-18T03:25:53Z to 2026-10-19T03:25:53Z	Hold at 0 while Fin is off. After Fin is on, 200 responses must rise above 0. All 503s after Fin is on is absent.	Sentry org sentry projects docs and develop-docs. Query has:http.method http.url:*/api/ask-ai* environment:production.

Regression watch

Watch search, the header Ask AI path, and server 500s on the same page tree. The layout wrap can fail every page. Search click handlers can fail Algolia. Do not treat gated /api/ask-ai/ 503s as a regression while Fin stays off.

Signal	Baseline	Window	Rule	Source
docs.search.query	484 queries in 24h. 517 in the prior 24h. Band 484-517. 408 of 484 had results. 7d sum is 12150, about 1736 per day.	2026-10-18T03:25:53Z to 2026-10-19T03:25:53Z, 2026-10-17T03:25:53Z to 2026-10-18T03:25:53Z, and 2026-10-12T03:25:53Z to 2026-10-19T03:25:53Z	Hold in the 484-517 weekend band. Weekday volume may rise toward the 7d mean. A collapse toward 0 is a regression.	Sentry metrics org sentry project docs. sum(value, docs.search.query) with environment:production.
Search and header client errors	6 events in 24h. 7 events in the prior 24h. Band 6-7. Share of search queries is 1.2-1.4%. Current titles are Algolia insights fetch failures and getItem on null.	2026-10-18T03:25:53Z to 2026-10-19T03:25:53Z and 2026-10-17T03:25:53Z to 2026-10-18T03:25:53Z	Hold in 6-7 per 24h. A rise above 15 events per 24h, or a new askAi stack, is a regression.	Sentry errors org sentry project docs. Query environment:production stack.filename:[*search/index.tsx*, *header.tsx*].
http.server 500 and 502	0 spans in 24h and in 48h. Status mix in 24h is 4418×200, 63×307, 33×404.	2026-10-18T03:25:53Z to 2026-10-19T03:25:53Z and 2026-10-17T03:25:53Z to 2026-10-19T03:25:53Z	Hold at 0 on paths other than /api/ask-ai/. Any new 500 or 502 is a regression.	Sentry traces org sentry project docs. Query span.op:http.server http.status_code:[500,502] environment:production.
LCP p75	420ms in 24h. 531ms over 48h. 865ms over 7d. Band 420-865ms.	2026-10-18T03:25:53Z to 2026-10-19T03:25:53Z, 2026-10-17T03:25:53Z to 2026-10-19T03:25:53Z, and 2026-10-12T03:25:53Z to 2026-10-19T03:25:53Z	Hold in 420-865ms. A rise above 900ms is a regression.	Sentry metrics org sentry project docs. p75(value, browser.web_vital.lcp) with environment:production.

Not observable

Ask AI open clicks are not in Sentry. Plausible Ask AI Referrer is not connected here. Cookie, storage, and third-party request absence cannot be checked from this telemetry. Answer quality, sanitization, and citations have no production signal. Signed webhook and event-export outcomes have no measured traffic. Datadog is not connected.

Get Slack DMs about this change: cursor.com/automations/rollouts

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want reviews to match your repository better? Bugbot Learning can learn team-specific rules from PR activity. A team admin can enable Learning in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit bc83804. Configure here.}